A Generative Approach to Simultaneous Diffeomorphic Registration and Lesion Segmentation of Neuroimages

Abstract

Image segmentation and image registration are two fundamental problems in computer vision and medical image processing. In image segmentation, one seeks to partition an image into meaningful segments by assigning a label to each pixel indicating which segment it belongs to. In image registration, one seeks to recover a spatial transformation that geometrically aligns two or more images, which allows downstream image analyses in which the registered images share a coordinate system. Image processing pipelines typically apply these procedures sequentially even though the segmentation of an image could improve its registration and registration of an image could improve its segmentation. With an appropriate parametrization, one can view these two tasks as an inference problem in which the spatial transformation and segmentation are latent variables. In this work, registration and segmentation are integrated through a hierarchical Bayesian generative framework. The framework models the data generating process of a set of magnetic resonance (MR) images of ischemic stroke lesioned brains. Under this framework, we simultaneously estimate a lesion tissue segmentation along with a spatial diffeomorphic transformation that maps a subject image into spatial correspondence with a healthy template image. The framework is evaluated on two-dimensional images both real and synthetic. Experimental results on real MR images show that simultaneous segmentation and registration can significantly improve the accuracy of lesion segmentation as well as the accuracy of registration near the lesion.